Continuous evidence for always-on agents常驻智能体需要怎样的持续证据
Release-time evidence cannot describe an agent that never has a release; what replaces it is still unsettled.
没有“发版时刻”的智能体无法靠发版证据来说明；替代方案尚无定论。
- Date · 日期
- 2026-10-03
- Status · 状态
- Draft — first pass; the central question is mostly open草稿 — 初稿；核心问题基本仍未解决
- Author · 作者
- 🌱
- Tags · 标签
- agents, evidence, monitoring, governance, always-on
- Question · 研究问题
- For an always-on agent with no release event, what is the minimum continuous evidence for a low-risk case, and who accepts it?对没有发版事件的常驻智能体，低风险场景下最少需要什么持续证据，由谁接受？
- Addresses · 对应
- Q6 (overlaps Q2, Q3) · H1
Executive summary摘要
Continuous monitoring is already normal in the rules for traditional models and high-risk AI: SR 26-2 expects ongoing monitoring and outcomes analysis, and the EU AI Act requires post-market monitoring. What is missing is a definition of how much continuous evidence is enough for a low-risk agent, and who accepts it. SR 26-2 explicitly leaves generative and agentic AI out of scope, so no US bank guidance answers this. I found no source that specifies a minimum for low-risk agents; the proposal below is my interpretation.
对传统模型和高风险 AI，持续监控本来就是规则的一部分：SR 26-2 要求持续监控与结果分析，欧盟《人工智能法案》要求上市后监测。缺少的是：低风险智能体“到底要多少”持续证据、由谁接受。SR 26-2 明确把生成式与智能体 AI 排除在适用范围之外，所以美国银行监管指引并未回答这个问题。我没有找到任何来源规定低风险智能体的最低标准；下文方案是我的解读。
Key findings主要发现
| # | Finding · 发现 | Cite |
|---|---|---|
| F1 | SR 26-2 describes ongoing model monitoring as checking whether a model still performs as expected as products, data or conditions change; frequency and scope depend on the model's nature, new data and materiality.SR 26-2 把持续监控描述为检查模型在产品、数据或环境变化后是否仍如预期；频率与范围取决于模型性质、新数据和重要性。 | [1] |
| F2 | SR 26-2 footnote 3 puts generative and agentic AI models out of scope, while saying a bank's own governance should still guide controls for them.SR 26-2 脚注 3 将生成式与智能体 AI 模型排除在范围外，同时说银行自身的治理实践仍应指导对它们的管控。 | [1] |
| F3 | SR 26-2 allows testing to be more limited for models with relatively less material impact: rigor scales with complexity and materiality.SR 26-2 允许对影响相对较小的模型采用范围更有限的测试：严格程度随复杂度与重要性而定。 | [1] |
| F4 | The EU AI Act requires a risk management process that is continuous over the lifecycle (Art. 9) and, for high-risk systems, a post-market monitoring system that actively collects performance data over the system's lifetime (Art. 72).欧盟《人工智能法案》要求在整个生命周期持续进行风险管理（第 9 条），并要求高风险系统建立上市后监测系统，在系统整个生命周期内主动收集性能数据（第 72 条）。 | [2] |
| F5 | Deployers of high-risk systems must keep automatically generated logs for at least six months, and assign human oversight to competent people (Art. 26).高风险系统的部署方须保存自动生成的日志至少六个月，并指派有能力的人员负责人工监督（第 26 条）。 | [2] |
| F6 | A study found GPT-4's accuracy on one prime-number task fell from 84% (March 2023) to 51% (June 2023), and its authors call for continuous monitoring of LLMs.一项研究发现 GPT-4 在某个质数判断任务上的准确率从 84%（2023 年 3 月）降至 51%（2023 年 6 月），作者呼吁对 LLM 持续监控。 | [3] |
| F7 | NVIDIA's OpenShell runtime enforces network, filesystem and process policy for agents, keeps credentials outside the workload, and records policy decisions in an OCSF audit trail; it lists Codex, Claude Code, Pi and Hermes as supported.NVIDIA 的 OpenShell 运行时对智能体强制执行网络、文件系统和进程策略，把凭据放在工作负载之外，并把策略决策记录到 OCSF 审计轨迹；列出支持 Codex、Claude Code、Pi 和 Hermes。 | [4] |
| F8 | OpenAI's own announcement (29 Sep 2026) describes dots as always-on agents with their own cloud computer that work in the background and, for some actions, wait for approval (an early tester's dot prepared an invoice and sent it after his approval). Its FAQ says sensitive actions need approval or human takeover, and that background “proactive research” uses read-only tools. Updated 2026-10-04 after reading OpenAI's pages; see the dots note.OpenAI 自己的发布页（2026-09-29）称 dots 是拥有独立云端电脑、在后台工作的常驻智能体，部分动作需等待批准（一位早期测试者的 dot 准备好发票，经其批准后发出）。其 FAQ 称敏感动作需批准或由人接手，后台“主动研究”只使用只读工具。已于 2026-10-04 在读过 OpenAI 页面后更新；见 dots 笔记。 | [6] [7] |
Evidence notes证据说明
SR 26-2 footnote 3: generative and agentic AI models "are not within the scope of this guidance". [1]
SR 26-2 脚注 3：生成式与智能体 AI 模型“不在本指引范围内”。[1]
The EU AI Act article text I read came from a mirror site (artificialintelligenceact.eu), not EUR-Lex directly. Its Art. 72 page states application from 2 Aug 2026, but the Digital Omnibus (secondary sources, not verified) reportedly moved high-risk obligations to 2 Dec 2027; check the dates against the Official Journal before citing them.
我读到的欧盟法案条文来自镜像站 artificialintelligenceact.eu，而非 EUR-Lex 原站。其第 72 条页面写明 2026-08-02 起适用，但据二手来源（未核实），Digital Omnibus 已将高风险义务推迟到 2027-12-02；引用前须对照官方公报核对日期。
Industry examples行业示例
| Company | Product | What it shows | Source |
|---|---|---|---|
| NVIDIA | OpenShell | Runtime policy enforcement plus OCSF audit trail: one candidate source of continuous action evidence.运行时策略强制加 OCSF 审计轨迹：持续动作证据的一个候选来源。 | [4] |
| OpenAI | Dots | The kind of agent without a release event that creates the problem (from OpenAI's own documents).产生这一问题的、没有发版事件的那类智能体（来自 OpenAI 自己的文档）。 | [6] [7] |
My interpretation我的解读
Existing rules already ask for continuous monitoring, so the novelty is not “monitor continuously” but “sized for a low-risk agent, accepted by whom”. The sources suggest the dimensions of a design (rigor scaled to materiality [F3]; human oversight by a competent person [F5]; logs as the raw evidence [F5, F7]) but no one has specified a low-risk tier for agents. A workable first design: (1) the permission and action log is the primary record; (2) automatic thresholds (permission violations, sampled-action error rate) trigger suspension; (3) a human accepts exceptions and periodic summaries rather than every run. Whether an independent agent could replace the human is Q3 and is untested here.
现有规则已经要求持续监控，所以新意不在“持续监控”，而在“按低风险智能体的规模设计、由谁接受”。来源提示了设计维度（严格程度随重要性调整 [F3]；由有能力的人员负责人工监督 [F5]；日志作为原始证据 [F5、F7]），但没有人为智能体规定过低风险层级。可行的初版设计：(1) 权限与动作日志是主要记录；(2) 自动阈值（越权次数、抽样动作错误率）触发暂停；(3) 人只接受例外和定期汇总，而非每次运行。独立智能体能否取代人是 Q3，本文未检验。
Hypotheses假设
For low-risk always-on agents, regulated firms will accept rule-based continuous gates with human review of exceptions before they accept agent-signed acceptance.
对低风险常驻智能体，受监管机构会先接受“规则型持续闸门 + 人审例外”，之后才会接受由智能体签字的验收。
Open questions未决问题
Which regulator, if any, will define a low-risk tier for agents? Where do action logs live (Q2)? Do the Dots-style approval prompts count as evidence of human oversight, or only as UX?
哪个监管机构会为智能体定义低风险层级？动作日志存放在哪里（Q2）？Dots 式的批准提示算人工监督的证据，还是只是产品交互？
What would change my view?什么会改变我的看法？
- Regulatory text that explicitly allows periodic (per-release) evidence for continuously running agents.
- 监管文本明确允许对持续运行的智能体采用定期（按版本）证据。
- Documented incidents showing per-release evaluation was sufficient for an always-on agent.
- 有记录的事故显示，对常驻智能体而言按版本评估已经足够。
Link to hypotheses与假设的关系
Weakly supports H1 only in the sense that evidence, not code, is what regulators and rules ask for. It does not test H1 directly. Not enough to change thesis confidence.
仅在“规则要求的是证据而非代码”这一意义上弱支持 H1；并未直接检验 H1。不足以改变论点置信度。
References参考文献
- Fed / OCC / FDIC, “SR 26-2: Revised Guidance on Model Risk Management,” 2026-04-17 (attachment PDF read: footnote 3, “Ongoing Model Monitoring”, validation and testing sections). https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm regulatory guidance · S-0006
- Regulation (EU) 2024/1689 (EU AI Act), Arts. 9, 26, 72, read via the mirror https://artificialintelligenceact.eu/article/72/ and /article/26/ (not EUR-Lex directly). regulation · S-0003
- Chen, Zaharia, Zou, “How is ChatGPT's behavior changing over time?” arXiv:2307.09009. https://arxiv.org/abs/2307.09009 paper · S-0011
- NVIDIA, “Add Runtime Controls to AI Agents with NVIDIA OpenShell,” 2026-09-28. https://developer.nvidia.com/blog/add-runtime-controls-to-ai-agents-with-nvidia-openshell/ vendor blog · S-0015
- SiliconANGLE, “OpenAI launches Dots, always-on AI agents in ChatGPT with their own cloud computers,” 2026-09-29 (secondary; OpenAI page returned 403). https://siliconangle.com/2026/09/29/openai-launches-dots-always-on-ai-agents-in-chatgpt-with-their-own-cloud-computers/ news · S-0016 Superseded for Dots facts by [6] and [7]; kept as the original lead.
- OpenAI, “Introducing dots,” 2026-09-29. https://openai.com/index/introducing-dots/ company announcement · S-0018
- OpenAI Help Center, “Dots privacy, security, and safety FAQs,” read 2026-10-04. https://help.openai.com/en/articles/20001529-dots-privacy-security-and-safety-faqs vendor documentation · S-0020
Last updated 2026-10-04 (F8 and Dots references now from OpenAI's own pages; the note's conclusions are unchanged).